6 of 800 posts
We had Jev score Elasticsearch hybrid search results and let a short Python policy do the reranking, taking nDCG@10 from 0.9351 to 0.9565 on 250 Amazon Shopping Queries, and the code is all here.
Idle indexing shards in Elasticsearch Serverless now drop their Lucene writers and segment readers from memory until the next write, freeing indexing-tier heap at the cost of a median 218ms wait on that first write.
We scored 6,000 ES|QL queries from four models against BIRD's answer key. Most misses come from mismatched join keys, SQL syntax the parser rejects, counting after a one-to-many join, or a value the model guessed.
Moving index builds to the GPU leaves the CPU free for queries, which is how vector indexing throughput went up 7x and p90 search latency fell 6x while indexing ran, with no change to recall.
Four agent performance questions your Agent Builder traces can answer, covering token spend by model, tool error rates, slow conversation turns, and recent prompts. The ES|QL for each is here, including the type cast SUM() needs.
Build AI agent orchestration where the workflow waits for a human approval and then executes the fix on its own, with nothing extra to provision and the whole decision trail queryable in Elasticsearch.
Blogsby Glyph